1
00:00:01,040 --> 00:00:01,670
All right.

2
00:00:01,680 --> 00:00:05,910
Welcome to a brand new module in this module.

3
00:00:05,940 --> 00:00:08,910
We're gonna be building a Caris model from scratch.

4
00:00:08,920 --> 00:00:11,070
Do some image classification.

5
00:00:11,160 --> 00:00:13,760
Previously we've been using a pre trained model.

6
00:00:13,770 --> 00:00:19,700
This time we're gonna train our own model and we're gonna do all of that from beginning to end.

7
00:00:19,710 --> 00:00:22,470
So we're gonna formulate our research question.

8
00:00:22,470 --> 00:00:29,100
We're going to gather our data clean our data and process our data explore and visualize some of the

9
00:00:29,100 --> 00:00:34,000
data and we're going to train our neural network on our local machine.

10
00:00:34,020 --> 00:00:40,320
And finally we're going to evaluate how we're doing a high actually want to kick off this module with

11
00:00:40,320 --> 00:00:44,550
a bit of a story you see in 2015.

12
00:00:44,550 --> 00:00:52,950
A Japanese man called Makoto Koichi kit was helping out his parents on their cucumber farm and one thing

13
00:00:52,950 --> 00:00:59,070
he realized was that his parents were sorting their cucumbers by the cucumbers quality.

14
00:00:59,130 --> 00:01:07,050
So for example a really fresh crispy and prickly cucumber sold at a huge premium to something that's

15
00:01:07,050 --> 00:01:13,740
like lower grade or damaged not all cucumbers are created equal apparently.

16
00:01:13,740 --> 00:01:21,420
However he also learned that farming cucumbers is a lot of work especially sorting cucumbers because

17
00:01:21,420 --> 00:01:28,520
apparently during harvest time his mother would spend about eight hours a day just sorting cucumbers.

18
00:01:28,710 --> 00:01:30,750
So that's when Makoto had an idea.

19
00:01:31,470 --> 00:01:38,510
What if he could use image classification to classify the cucumbers according to their grade.

20
00:01:38,520 --> 00:01:45,720
So like premium normal or lower grade hen then he could put those cucumbers on a conveyor belt and probably

21
00:01:45,720 --> 00:01:47,790
drop them off in the appropriate box.

22
00:01:48,540 --> 00:01:56,060
So this was the task that Makoto set himself the first step was gathering lots of data on his cucumbers.

23
00:01:56,070 --> 00:02:01,290
He started taking pictures of them and he started labelling these pictures of the cucumbers according

24
00:02:01,290 --> 00:02:03,210
to their grade hand.

25
00:02:03,270 --> 00:02:11,100
Makoto spent about three months or so taking about 7000 pictures of these cucumbers and this data set

26
00:02:11,160 --> 00:02:13,470
was what he used to train his algorithm.

27
00:02:13,470 --> 00:02:20,370
First he downsized all these pictures into like 80 pixel by 80 pixel images and then he used to tensor

28
00:02:20,400 --> 00:02:27,360
flow on his own computer to train a neural network and then his computer spent about two days training

29
00:02:27,360 --> 00:02:29,960
the model and calculating the weights.

30
00:02:31,160 --> 00:02:37,100
And after all that was done all he had to do was build some sort of sorting machine which had the cucumbers

31
00:02:37,100 --> 00:02:38,810
coming on a conveyor belt.

32
00:02:38,810 --> 00:02:44,390
They would make a stop in a photo booth where a picture was taken of them and then they moved down the

33
00:02:44,390 --> 00:02:48,770
conveyor and were deposited in the appropriate box.

34
00:02:48,770 --> 00:02:55,790
So this is a real story of a real person using image classification to solve a real problem.

35
00:02:55,790 --> 00:02:58,810
And we're going to take some inspiration from this story.

36
00:02:58,850 --> 00:03:01,210
We're also going to build our own image classifier.

37
00:03:01,470 --> 00:03:04,190
And it's going to classify 10 different things.

38
00:03:04,190 --> 00:03:16,870
It's going to differentiate among dogs ships horses frogs deer cats trucks cars planes and birds.

39
00:03:16,940 --> 00:03:21,280
In total we're actually going to have 10 different classes.

40
00:03:21,320 --> 00:03:25,810
That's the word that you'll see in the literature for the different categories.

41
00:03:26,800 --> 00:03:31,840
And the simple model that we're going to build is going to be an artificial neural network called a

42
00:03:31,890 --> 00:03:34,070
multi-layer perception.

43
00:03:34,090 --> 00:03:36,050
That's the name for it.

44
00:03:36,100 --> 00:03:40,850
It's one of the simplest kinds of neural networks and we're going to see how it fares.

45
00:03:40,960 --> 00:03:46,070
We're gonna see what kind of challenges we come up against and how we can resolve them.

46
00:03:46,090 --> 00:03:48,840
So I hope you're as excited as I am.

47
00:03:48,940 --> 00:03:52,300
I can't wait to see you in the next lesson where we're gonna do the setup.